Periodic Reporting for period 1 - AI-PROGNOSIS (Artificial intelligence-based Parkinson’s disease risk assessment and prognosis)
Okres sprawozdawczy: 2023-07-01 do 2024-12-31
Parkinson's disease (PD) is the most common progressive movement disorder, with diverse motor and non-motor symptoms and no cure. PD is often missed or misdiagnosed as early signs are subtle and common with other diseases, while selecting the optimal treatment is usually a lengthy process, leading to unnecessary suffering. AI-PROGNOSIS (www.ai-prognosis.eu) aims to advance PD diagnosis and care by harnessing the power of artificial intelligence (AI). The project will develop predictive models for PD risk assessment and prognosis (in terms of progression and response of patients to medication) based on in-depth health and genetic data and combine them with digital biomarkers of key disease symptoms and determinants measured in everyday life. Project work will lead to the creation of the AI-PROGNOSIS ecosystem of apps, a digital health toolkit offering persons affected by PD and healthcare professionals (HCPs) valuable insights for informed health management.
The first period of the project laid the foundation and saw the research and development (R&D) phase unfold.
Foundation: A broad review of recent advances in PD research and technologies was carried out that explored markers, determinants, and AI-based approaches for the prediction of PD risk, progression, and medication response, as well as approaches for tracking motor and non-motor symptoms via digital means, such as wearables. Datasets of value to the project, including the Parkinson's Progression Markers Initiative (PPMI) and the Critical Path for Parkinson’s (CPP) databases, were identified and accessed, while a data harmonisation strategy was put in place, based on the OMOP common data model. To steer R&D, a Trustworthy AI development framework was created to guide the development of the project’s AI-enabled tools.
User research and co-creation: To ensure sustainable stakeholder engagement, a patient panel was formed and user research was conducted with the panel and HCPs. This research identified key user needs and attitudes toward the project’s envisioned digital health tools, the mAI-Health and mAI-Care patient-facing apps for PD screening and management, respectively, and the mAI-Insights decision-support platform for HCPs. Co-creation activities followed, including workshops to specify user requirements for trusting the AI systems to be embedded in the AI-PROGNOSIS tools and user experience co-design sessions for the mAI-Care app with the patient panel.
Data analysis and AI: The first R&D phase focused on harmonising a subset of the PPMI dataset according to the OMOP model, developing digital assessments of key PD symptoms and determinants, and creating the early versions of the AI models for assessing PD risk and predicting disease progression and medication response. Digital biomarkers were developed to passively track REM sleep behaviour disorder (RBD), daytime sleepiness, rest tremor, slowness of movement, physical activity, and fine motor skills through various data sources, including wrist motion data and keystroke timings during touchscreen typing. Active assessments were also created, including mobile versions of standardised cognitive tests and a prototype motor function test using smartphone camera-based body tracking to assess posture, balance, and walking. For PD risk, genetic variations linked to the disease were explored in the AMP-PD cohorts, while a machine learning model to detect PD was developed based on smartwatch-captured daily mobility and sleep characteristics of individuals in the Verily Study Watch cohort. For disease prognosis, early versions of a predictive model of the UPDRS Part III score - a clinical score of motor impairement - were developed based on PPMI patients’ phenotypic data, while models predicting dyskinesias, a common side effect of PD medication, were developed based on the CPP cohorts’ clinico-demographic and medication data.
Digital health ecosystem: Mapping the identified user needs to features of the tools and detailing the technical specifications and high-level architecture of the AI-PROGNOSIS system laid the foundation for software development. Α secure Cloud-based data management system was set up, monitoring mechanisms were put in place, and a study app was developed, implementing the core smartphone and smartwatch data collection and serving the needs of the dBM-DEV study, the project’s first clinical study. Based on this, early versions of the mAI-Health and mAI-Care patient-facing apps along with an initial version of the HCPs’ mAI-Insights platform were developed.
Clinical studies: The period saw the initiation of the dBM-DEV study (NCT06444789), a multi-centre study collecting clinico-demographic, wearable, and smartphone data from persons with RBD and PD to extend and validate the digital biomarkers of PD symptoms and progression determinants, being implemented by the project. Steps were also taken to ensure cross-site feasibility and appropriate outcomes for the two multi-centre proof-of-concept studies, AI-PRA and AI-PMP, which aim to validate the predictive models for PD risk, progression, and medication response, and evaluate the acceptance of the associated AI-PROGNOSIS tools.
Foundation: A broad review of recent advances in PD research and technologies was carried out that explored markers, determinants, and AI-based approaches for the prediction of PD risk, progression, and medication response, as well as approaches for tracking motor and non-motor symptoms via digital means, such as wearables. Datasets of value to the project, including the Parkinson's Progression Markers Initiative (PPMI) and the Critical Path for Parkinson’s (CPP) databases, were identified and accessed, while a data harmonisation strategy was put in place, based on the OMOP common data model. To steer R&D, a Trustworthy AI development framework was created to guide the development of the project’s AI-enabled tools.
User research and co-creation: To ensure sustainable stakeholder engagement, a patient panel was formed and user research was conducted with the panel and HCPs. This research identified key user needs and attitudes toward the project’s envisioned digital health tools, the mAI-Health and mAI-Care patient-facing apps for PD screening and management, respectively, and the mAI-Insights decision-support platform for HCPs. Co-creation activities followed, including workshops to specify user requirements for trusting the AI systems to be embedded in the AI-PROGNOSIS tools and user experience co-design sessions for the mAI-Care app with the patient panel.
Data analysis and AI: The first R&D phase focused on harmonising a subset of the PPMI dataset according to the OMOP model, developing digital assessments of key PD symptoms and determinants, and creating the early versions of the AI models for assessing PD risk and predicting disease progression and medication response. Digital biomarkers were developed to passively track REM sleep behaviour disorder (RBD), daytime sleepiness, rest tremor, slowness of movement, physical activity, and fine motor skills through various data sources, including wrist motion data and keystroke timings during touchscreen typing. Active assessments were also created, including mobile versions of standardised cognitive tests and a prototype motor function test using smartphone camera-based body tracking to assess posture, balance, and walking. For PD risk, genetic variations linked to the disease were explored in the AMP-PD cohorts, while a machine learning model to detect PD was developed based on smartwatch-captured daily mobility and sleep characteristics of individuals in the Verily Study Watch cohort. For disease prognosis, early versions of a predictive model of the UPDRS Part III score - a clinical score of motor impairement - were developed based on PPMI patients’ phenotypic data, while models predicting dyskinesias, a common side effect of PD medication, were developed based on the CPP cohorts’ clinico-demographic and medication data.
Digital health ecosystem: Mapping the identified user needs to features of the tools and detailing the technical specifications and high-level architecture of the AI-PROGNOSIS system laid the foundation for software development. Α secure Cloud-based data management system was set up, monitoring mechanisms were put in place, and a study app was developed, implementing the core smartphone and smartwatch data collection and serving the needs of the dBM-DEV study, the project’s first clinical study. Based on this, early versions of the mAI-Health and mAI-Care patient-facing apps along with an initial version of the HCPs’ mAI-Insights platform were developed.
Clinical studies: The period saw the initiation of the dBM-DEV study (NCT06444789), a multi-centre study collecting clinico-demographic, wearable, and smartphone data from persons with RBD and PD to extend and validate the digital biomarkers of PD symptoms and progression determinants, being implemented by the project. Steps were also taken to ensure cross-site feasibility and appropriate outcomes for the two multi-centre proof-of-concept studies, AI-PRA and AI-PMP, which aim to validate the predictive models for PD risk, progression, and medication response, and evaluate the acceptance of the associated AI-PROGNOSIS tools.
The project has produced results that add to the state-of-the-art and have impact potential. The domain review and novel user research findings on AI-enabled solutions for screening and care can guide future innovations, while the Trustworthy AI development framework further provides a blueprint for creating reliable AI systems. Harmonising a core part of PPMI with OMOP enables unification with similar OMOP data and facilitates cross-cohort research. Early version of digital biomarkers of RBD and daytime sleepiness are paving the way in the nascent field of remote assessment of sleep disturbances in PD, while the implementation of existing algorithms for rest tremor and slowness of movement assessment with new datasets, created valuable evidence for their robustness and reuse. The first version of the PD risk assessment model and the early validation evidence, based on the US-originating Verily Study Watch dataset, is one of the first attempts to PD detection based on longitudinal mobility and sleep data, and further adds to similar, recently-published work that relied on UK Biobank data. Lastly, the ongoing dBM-DEV study is collecting a comprehensive dataset that can support the development and validation of digital assessments for PD symptoms.